An improved insulator self-explosion detection method based on group-level pruning for the YOLOv7-tiny algorithm

With the construction of intelligent grids, unmanned aerial vehicle have been widely employed to inspect transmission lines. The inspection process generates a large amount of data, which requires a lightweight model to reduce computational overhead. Here, we propose an improved model based on YOLOv7-tiny with group-level pruning to reduce the model size, which achieves a balance between detection accuracy and speed. Firstly, we replace the activation function with the Funnel activation function to optimize the activation domain dynamically. Second, we introduce a lightweight DFC attention mechanism to enhance the ability of backbone to extract long-range features. Finally, we use adaptively spatial feature fusion network to reduce semantic degradation during feature fusion. We group the parameters according to their dependencies and use a consistent sparse approach to obtain parameter importance. The redundant parameter groups were pruned to achieve model light-weighting. Experimental results show that the improved model achieves 95.6% detection accuracy after pruning. Compared with YOLOv7-tiny, the computational complexity is reduced by 53% and the processing speed is increased by 48.1% to 73 frames per second.